- Implemented a deep CNN using the DeXpression architecture for automatic facial expression recognition (anger, fear, surprise, etc.).
- Trained and evaluated the model with 5-fold Cross-Validation on the Extended Cohn–Kanade (CK+) dataset. Used data preprocessing, and regularization to improve robustness and reduce overfitting.
- Achieved 99% mean training accuracy and 98% mean test accuracy, demonstrating robust generalization on canonical FER benchmarks and matching state-of-the-art performance on CK+.
- Repeated the same with Transformer architecture.
Architecture:
DeXpression
- (stacked convolution + pooling + maybe inception-like blocks, final FC + softmax).
Preprocessing ?
- Grayscale or RGB?
- Cropping / alignment?
- Normalization?
- Data augmentation?
5-fold CV ? Generalization to unseen data.
Overfitting checks and monitors (train vs. val curves, early stopping, regularization).
Classes Emotions, # of classes, notable confusions (e.g., fear vs surprise).